lkml.org 
[lkml]   [2021]   [Feb]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v5 05/10] userfaultfd: add minor fault registration mode
On Fri, Feb 12, 2021 at 02:51:17PM -0800, Axel Rasmussen wrote:
> On Fri, Feb 12, 2021 at 2:44 PM Peter Xu <peterx@redhat.com> wrote:
> >
> > On Fri, Feb 12, 2021 at 10:21:45PM +0000, Matthew Wilcox wrote:
> > > On Thu, Feb 11, 2021 at 11:28:09AM -0800, Axel Rasmussen wrote:
> > > > Ah, I had added this just after VM_UFFD_WP, without noticing that this
> > > > would be sharing a bit with VM_LOCKED. That seems like not such a
> > > > great idea.
> > > >
> > > > I don't see another unused bit, and I don't see some other obvious
> > > > candidate to share with. So, the solution that comes to mind is
> > >
> > > it'd be even better if you didn't use the last unused bit for UFFD_WP.
> > > not sure how feasible that is, but you can see we're really short on
> > > bits here.
> >
> > UFFD_WP is used now for anonymouse already.. And the support for hugetlbfs and
> > shmem is in rfc stage on the list.
> >
> > Is it possible to use CONFIG_ARCH_USES_HIGH_VMA_FLAGS here? So far uffd-wp is
> > only working for 64 bit x86 too due to enlarged pte space. Maybe we can also
> > let minor mode to only support 64 bit hosts.
>
> At least for my / Google's purposes, I don't care about 32-bit support
> for this feature. I do care about both x86_64 and arm64, though. So
> it's a possibility.
>
> Alternatively, the "it's an API feature not a registration mode"
> approach I sent in my v6 also works for me, although it has some
> drawbacks.

Per-vma has finer granularity and logically more flexible. If it's low hanging
fruit, let's think about it more before giving up so quickly.

Sorry I commented late for this - I got diverged a bit in the past days. While
you worked on it so fast (which in many cases still a good thing :).

>
> Another option is, would it be terrible to add an extra u16 or u32 for
> UFFD flags to vm_area_struct (say within vm_userfaultfd_ctx)?
> Historically we've already added a pointer, so maybe an extra say 16
> bits isn't so bad? This would avoid using *any* VM_* flags for UFFD,
> even VM_UFFD_MISSING could be in this new flag field.

For 64bit hosts there're still places for vm_flags. It's just 32bit, while
there's option to make it 64bit-only. Even if we'd add a new field, those bits
were still unused on 64bit hosts. IMHO we should try to use them before adding
new field which will actually impact all hosts.

Thanks,

--
Peter Xu

\
 
 \ /
  Last update: 2021-02-13 00:05    [W:0.226 / U:0.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site